The fifth international conference on regional climate (ICRC 2023), organised by World Climate Research Programme’s (WCRP) coordinated downscaling experiment (CORDEX), has just completed. It was a hybrid on-site/online conference with hubs in both Trieste/Italy (hosted by the International Centre on Theoretical Physics, ICTP) and Pune/India.
The hybrid set-up, with video links between the two hubs and digital attendence through zoom, was a change from previous ICRCs held in ICTP (2011), Brussels (2013), Stockholm (2016), and Beijing (2019). It worked impressively well, and the CORDEX ICRC 2023 streaming is available from the WCRP CORDEX YouTube channel.
It seems as an eternity since the previous ICRC before the COVID pandemic, so I was curious to see how things have progressed since then. It was also interesting to compare my impressions from this conference with my blog posts here on RealClimate from the first ICRC in Trieste, the second in Brussels, the third ICRC in Stockholm, I see that questions concerning uncertainty and added value are still being debated.
Developments within regional climate modelling
A new development, however, is the coupling of regional climate models (RCM), which traditionally have been limited to a volume of the atmosphere with some input from the surface. There were presentations on attempts to couple the RCMs to the sea, hydrology, and glaciers glacier. In this kind of coupling there is a mutual exchange of energy and mass, and hence involves a two-way interaction.
Another focus was on so-called “convective-permitting”, which are RCMs with higher spatial resolution (the size of grid boxes are a few kms) so that they explicitly can calculate convective processes, such as vertical ascent, cloud formation and precipitation. There is also work on RCMs so that they better account for aerosols and how they vary over time and space.
Artificial intelligence and machine learning
Since the previous ICRCs, artificial Intelligence (AI) has become a hotter topic, also within empirical-statistical downscaling (ESD). However, deep learning algorithms, such as artificial neural nets, were already discussed in 1999 (e.g. Zorita and von Storch, 1999), but the conclusion in 1999 was that the more advanced deep learning algorithms didn’t perform better than simpler methods such as analogs.
I’m still not convinced that AI and deep learning will take us much further now, even if the tests from a number of attempts may look impressive at first sight (how many trials have there been and how many of them failed?).
One problem is AI’s reliance on large data volumes. The data used to train AI often includes a range of different variables which may move in different directions in the future when the world heats up.
Another problem is that different global climate models (GCMs) have different biases and quirks, and that may cause problems when transferring AI trained on historical observations to the model world (or from one model to another).
Provision of climate information to society
One session was devoted to the interaction with society. To me, it seems that CORDEX is not yet ready to provide society with robust and actionable information, despite the initiative from the WCRP called regional information for society (RIfS).
RIfS has taken a long time since it was conceived in 2020 (my impression is that it’s still not ready). There is an urgency underscored by the numerous reports of weather-related calamities around the world – we are not even adapted to the current climate and a newsreport from Washington Post reveals that 2023 September month global mean temperature was probably about 1.7°C above the preindustrial level.
There is a WCRP Open Science Conference (OSC 2023) in Kigali October 23-27 where RIfS probably will be discussed further, but it’s not clear to me what it’s all about or who is involved. My impression is that the WCRP and RIfS have closer links to more academic university communities than for instance more applied and operational national meteorological services.
Many national meteorological services have already established routines and are experienced in providing regional weather and climate information to society.
For instance, the Norwegian Meteorological Institute collaborates with various institutions and authorities, such as the Norwegian Water Resources and Energy Directorate (NVE), the Norwegian Institute of Public Health (NIPH), the Norwegian Directorate for Civil Protection, power production (StatKraft) and grid (Statnett), road authorities, aviation, rail, and defense. Our experience is that relevant information flows quite well within such a professional network.
Climate services in Norway differ from weather forecasting as they aim at the municipalities and need to reach local engineers and policymakers. There are other hurdles that need to be overcome when establishing new routines compared to state authorities.
For instance, small municipalities typically lack resources, the know how and incentives. They often have set protocols and routines that are not designed to accommodate climate change adaptation.
Typical topics include water management and area planning. A unique approach in Norway is to channel climate information through the trade union for civil engineers, Tekna (which is both a professional society and a trade union), e.g. through professional courses.
The meteorological service also has some experience with impact studies and we have collaborated on e.g. toxicology (SETAC), health, biology, national heritage, indigenous people (reindeer herders), and disaster risk reduction (e.g. flooding, earth-slides). So even if the progress is slow with RIfS, there is plenty of activities on applied research relevant for society.
Some self criticism
A remark made during the ICRC made me wonder about the question: Would our regional climate modelling community benefit from more critical reflections and discussion about what we should avoid?
My impression is that there are some cases of flawed use of downscaling that we don’t call out often enough. Nevertheless, it’s necessary to be extra critical and quizzical when our results are used for climate change adaptation and impact studies in order to avoid maladaptation and misleading impact studies.
One message that I tried to remind my colleagues is that everybody, who downscales global climate model output for use as regional and local climate projections, must read Deser et al. (2012) and account for random regional climate variations on scales up to decades.
There are too many examples of regional and local climate projections based on one or a small number of global climate model simulations. The “law of small numbers” implies a minimum number of independent simulations in an ensemble (Rabin, 2002).
If one picks the results from one climate model one gets a projection, but if one were to chose another computation from the very same model, then the projection would look quite different.
The difference is due to the chaotic nature of natural regional climate variations. But if we estimate statistics (e.g mean or probability distribution/histogram) on say 100 simulations, then such statistics won’t be much affected if we were to change one or some of the model simulations. This is what we mean by robust results and it’s also useful that statistical properties often are more predictable than individual outcomes.
There are some fundamental principles that received little attention during the ICRC, including the models’ minimum skillful scales (Takayabu et al., 2016) and the evaluation of the global climate models’ (which provide the boundary conditions for downscaling) ability to skillfully reproduce large-scale change.
Regional climate modelling and world weather attribution
Another issue that puzzled me is the concept of world weather attribution – how does it fit into the limitations and the divergence highlighted here at the ICRC in connection to downscaling?
And how well do global climate models, limited by their minimum skillful scales, represent specific events such as torrential rainfall and heatwaves. There are still some limitations concerning blocking-frequency and the representation of convection (Schumacher et al. 2023).
Enhancing the signal-to-noise ratio during conferences
One trait that I find typical of many conferences, such as the ICRC, is that too much details are crammed into 10 minutes. I suffered from information overload and I’m sure I missed many points.
Perhaps one trick for enhancing our chance of really reaching our audience is try to keep in mind that when we present our results, each one of us is just one in a hundred presenters. Hence text on the slides should be kept to less than a minimum – no more than just a few words.
It goes wrong when we try to read the text on the slides while we also try to listen to the presenter talking – two competing brain activities.
A test I try myself is to try to recall how much do I remember the day after I returned from a conference, and if I remember some of the talks, I try to explain why I remembered them.
It’s also useful to keep in mind that the purpose of a conference presentation is not to show how clever we are or share all our results, but make the attendees so interested in our work that they will contact us later or read up on our publications.
The past and a possible future for downscaling
This ICRC, and looking back on the previous ones, made me ponder about the future for downscaling. There are some concerns that RCMs will become obsolete when the spatial resolution in GCMs are increased.
ESD, on the other hand, may be less affected as it may be used to estimate statistics not directly connected to the atmospheric physics, e.g. biological/health statistics or number of certain events. Also, ESD is useful for downscaling statistical properties (e.g. parametres of probability density functions) directly, something that cannot be done with a high-resolution GCM.
I also expect regional climate information in time will become an increasingly hot topic in global climate summits (COPs), especially when representatives and negotiators from various nations want to know what consequences other nations are likely to face. This will give them extra information about what issues are more important for their opponents.
ICRC 2023 paid a tribute to Filippo Giorgi who retries next year. He has been a titan within regional climate modelling and had an honourable career.
References
- C. Deser, R. Knutti, S. Solomon, and A.S. Phillips, “Communication of the role of natural variability in future North American climate”, Nature Climate Change, vol. 2, pp. 775-779, 2012. http://dx.doi.org/10.1038/nclimate1562
- M. Rabin, “Inference by Believers in the Law of Small Numbers”, The Quarterly Journal of Economics, vol. 117, pp. 775-816, 2002. http://dx.doi.org/10.1162/003355302760193896
- I. TAKAYABU, H. KANAMARU, K. DAIRAKU, R. BENESTAD, H.V. STORCH, and J.H. CHRISTENSEN, “Reconsidering the Quality and Utility of Downscaling”, Journal of the Meteorological Society of Japan. Ser. II, vol. 94A, pp. 31-45, 2016. http://dx.doi.org/10.2151/jmsj.2015-042
- D. Schumacher, J. Singh, M. Hauser, E. Fischer, and S. Seneviratne, “Why climate models underestimate the exacerbated warming in Western Europe”, 2023. http://dx.doi.org/10.21203/rs.3.rs-3314992/v1
